Lightweight RT-DETR for Steel Surface Defect Detection

To address the issues of computational redundancy, insufficient real-time performance, and low recognition rates in steel surface defect detection using the RT-DETR architecture, this paper proposes a lightweight real-time defect detection model (RT-DETR) integrating Ghost convolution and Convolutional Block Attention Module (CBAM).Through dual optimization strategies: Ghost convolution replaces traditional convolutional layers to generate feature maps via linear transformations, compressing parameters to 42.37M and reducing computation to 128.3 GFLOPS; the CBAM module is embedded in the feature fusion network to enhance separability of micro-scale defects in complex backgrounds through channel weight calibration and spatial pixel focusing; a multi-scale feature pyramid fusion mechanism solves feature degradation during hierarchical transmission. Experiments on the NEU-DET benchmark show this method achieves real-time processing at 46 FPS and 0.81 mAP50 comprehensive detection accuracy for six defect types, significantly outperforming mainstream models like YOLOv10n with state-of-the-art results across defect categories. The model effectively resolves traditional methods' limitations of low sensitivity to irregular defect boundaries and inadequate recognition of rolled-in scale defects, providing a deployable solution for high-precision industrial inspection systems.